Ollama v0.32.6: MLX Speculative Decoding Auto-Enabled + Streaming Format Corrected

Ollama v0.32.6 (Aug 5) automatically enables MLX speculative decoding for Apple Silicon and corrects the /v1/chat/completions streaming format to match OpenAI's wire protocol exactly.

MLX speculative decoding is now automatic. The MLX engine uses a model's Multi-Token Prediction (MTP) head for speculative decoding without any configuration. Qwen3.5 and other MTP-capable models run noticeably faster on M-series chips — no flag to set, no separate draft model to manage. If the model doesn't support speculative decoding, Ollama falls back silently to standard inference.

OpenAI streaming wire format corrected. The /v1/chat/completions SSE stream now behaves like the real OpenAI API: role appears only in the first chunk, finish_reason arrives on its own trailing chunk, and token usage arrives in a separate chunk when stream_options.include_usage is set. If you wrote streaming parsers that worked around Ollama's previous format quirks, test them after upgrading.

Why it matters: Speculative decoding is a free latency improvement for local inference with no accuracy tradeoff, and the streaming fix removes the last common friction point when substituting Ollama for the OpenAI SDK in existing code.

Sources: Ollama v0.32.6 release — GitHub · Ollama August 2026 updates — Releasebot